This is a free quality signal we can compute with no human label and no ground truth, layered on top of the confidence score from test 0.3. It won't replace confidence, but it catches a different kind of mistake: a confident, smooth line that's simply tracing the wrong thing on the sail.
| Measure | Value |
|---|---|
| Wiggle-vs-error correlation (Spearman) | 0.62 |
| Detections checked | 1091 |
| "Batten-grab" lines caught (smooth line, far from any real stripe) | 15 of 1286 (1.17%) |
We fit a smooth curve to every line the model drew (no human label needed) and measured how far the real points wander from that fit — "wiggle." Lines that are perfectly smooth but land far from any real stripe are almost always the model latching onto a batten instead of a stripe.
Source: 332-photo holdout, frozen 2026-08-25. 1,286 detections, all with valid curve fits.